Background technology
The precision of robot is reflection robot performance's a important indicator, and it comprises absolute fix precision and repetitive positioning accuracy.The absolute fix trueness error is the deviation between robot actual motion and the desired motion, by determinacy initial error (like connecting rod parameter error, pair clearance etc.) generation; The repetitive positioning accuracy error is a robot when repeating same desired motion, the mutual dispersion degree between the actual motion of robot, by randomness initial error (like joint servo positioning error etc.) generation.
Usually, the general industrial robot has higher repetitive positioning accuracy, yet its absolute fix precision is poor, and repetitive positioning accuracy reaches the robot of 0.1mm, and its absolute fix trueness error but can have 2-3mm.Teaching has utilized the repetitive positioning accuracy of robot can reach higher absolute fix precision, but in practical application, a lot of occasions are very high to the absolute fix accuracy requirement of industrial robot, can not accomplish through the mode of teaching again usually.So, the absolute fix precision of robot compensated just seem most important.
Document " Xia Kai, Chen Chongduan, turbulent waves, etc. the neural network [J] of compensation robot positioning error. robot, 1995,17 (3): 171-176. " in order to improve the bearing accuracy of robot, the neural net method of compensation robot positioning error has been proposed.Article is to the RM-501 robot with five degrees of freedom, through the error that dual mode is located with multilayer perceptron neural networks compensate robot kinematics equation, is based on the neural networks compensate of joint coordinates respectively and based on the neural networks compensate of rectangular coordinate position.The core of this method is to utilize artificial neural network to have very strong self study, adaptive ability; Obtain the action rule in robot kinematics's geometric parameter error, non-geometric parameter error equal error source through training; Expectation joint angle or rectangular coordinate are compensated, thereby improve the absolute fix precision of robot.This method has been avoided loaded down with trivial details modeling of other traditional scaling methods and parameter identification process, but not enough below in practical application, existing:
1) for make network after the training can reach certain precision and adapt in the robot envelope have a few, neural network training needs a large amount of learning samples, so the surveying work amount is big;
2) this method need convert rectangular coordinate to joint coordinates, therefore need invert to robot kinematics's equation and separate, and this process calculated amount is big, and effect is undesirable near singular point;
3) test findings shows that robot is still not ideal enough in calibrated bearing accuracy.
Summary of the invention
The present invention to the deficiency of prior art existence, and proposes a kind of industrial robot space lattice precision compensation method based on the particle group optimizing neural network for improving the absolute fix precision of industrial robot.
This method comprises the steps:
Step 1: in the envelope scope of industrial robot, be divided into a series of cube grid to the whole enveloping space by certain step-length;
Step 2: measure and set up basis coordinates system of robot through laser tracker; Theoretical coordinate with eight summits of each cube grid of division in the step 1 under several different environment temperature levels comes control robot to position, and measures and record actual location coordinate with laser tracker;
Step 3: set up BP neural network model, and train with the data that collect in the step 2 based on particle group optimizing;
Step 4: (X, Y is Z) with residing actual environment temperature T for the target localization coordinate of expecting to arrive any point P in the envelope scope
1) searches the cube grid that this P belongs to;
2) theoretical coordinate and these four parameters of environment temperature on eight summits that respectively P belonged to the cube grid are as the input of neural network, thereby dope the actual location coordinate on corresponding eight summits;
3) calculate this P and eight summit K that belong to the cube grid
iThe actual location coordinate apart from d
i, with calculate apart from d
iCarry out anti-distance weighted eight the summit K that try to achieve
iWeights q with respect to this P
i, i=1 wherein, 2 ... 8, down with;
4) with the weights q that tries to achieve
iCome eight summit K
iThree directions of X, Y, Z on positioning error carry out space interpolation respectively, budget goes out the error on three directions of this P;
5) (X, Y Z) carry out inverse modified, accomplish the bearing accuracy compensation of robot at this P to the theoretical coordinate of this P with the error of trying to achieve.
The present invention has following technique effect:
1) to the industrial robot of different model,, can reduce the workload of measurement effectively, help industrial robot and drop into application apace through confirming to divide the maximum step-length of grid.
2) this method is carried out in cartesian coordinate system, compares normal solution and contrary the separating that need not carry out the robot kinematics with common robot scaling method, and computation process is simply rapid, can be implemented in the line compensation.
3) take all factors into consideration the error that robot motion's mathematic(al) parameter, load and variation of ambient temperature are brought, significantly improved the absolute fix precision of robot, made calibrated industrial robot can adapt to application scenario widely.
Embodiment
The step of industrial robot space lattice precision compensation method that the present invention is based on the particle group optimizing neural network is following:
Step 1: in the envelope scope of industrial robot, become a series of cube grid by the area dividing to be processed of certain step-length in the robot enveloping space;
Step 2: measure and set up basis coordinates system of robot through laser tracker; Theoretical coordinate with eight summits of each cube grid of division in the step 1 under several different temperature levels comes control robot to position, and measures and record actual location coordinate with laser tracker;
Set up that related step is between laser tracker and the robot basis coordinates system:
1) spherical fixed reflector SMR is fixed on the TCP of end effector, and keeps A2 to immobilize,, measure a series of points that are positioned on the circumference with the FARO laser tracker through rotation A1 axle to the position (angle) of A6 axle;
2) the CAM2 Measure software that utilizes FARO to carry, the measurement point that step 1 is obtained simulates a plane and a circle successively, thereby obtains the theoretical coordinate in the center of circle;
3) series of points on the robot measurement base plane and simulate a plane is done the skew that side-play amount is the SMR radius to this plane again;
4) project to the center of circle that obtains in the step 2 in the plane that step 3 skew obtains, obtain the origin position of robot;
The mounting hole of two horizontal symmetrical on robot ring flange when 5) robot measurement is in null position is calculated the mid point of these two measurement points then, and the offset planes that in step 3, obtains the gained mid point is again done projection, obtain being on the X axle a bit;
6) begin zero point to do projection to this offset planes that in step 3, obtains again from the man-machine tool of machine by counterclockwise rotating A1 axle (less than the arbitrary value of 90 degree) and measuring, be in+on the X Y plane a bit;
7) 2 that utilize initial point and step 5 and 6 to obtain construct coordinate system, and this coordinate system is the robot coordinate system.
Step 3: for expectation arrive any point P in the envelope scope theoretical coordinate (X, Y, Z) with environment temperature T,
1) searches the cube grid that this P belongs to;
2) theoretical coordinate and these four parameters of environment temperature on eight summits that respectively P belonged to the cube grid are as the input of neural network, thereby dope the actual location coordinate on corresponding eight summits;
3) calculate this P and eight summit K that belong to the cube grid
iThe actual location coordinate apart from d
i, with calculate apart from d
iCarry out anti-distance weighted eight the summit K that try to achieve
iWeights q with respect to this P
i, i=1 wherein, 2 ... 8, down with;
4) with the weights q that tries to achieve
iCome eight summit K
iThree directions of X, Y, Z on positioning error carry out space interpolation respectively, budget goes out the error on three directions of this P;
5) (X, Y Z) carry out inverse modified, accomplish the bearing accuracy compensation of robot at this P to the theoretical coordinate of this P with the error of trying to achieve.
In the said step 3:
In the step 4:
(3)
In the step 5:
X
Repair=X+ Δ X; Y
Repair=Y+ Δ Y; Z
Repair=Z+ Δ Z
(4)
In the formula: (X, Y Z) are the theoretical coordinate of a P; (X
i', Y
i', Z
i') be the actual location coordinate that comprises eight summits of cube grid of a P; (Δ X
i, Δ Y
i, Δ Z
i) be actual coordinate and the theoretical coordinate deviation that comprises eight summits of cube grid of a P; (Δ X, Δ Y, Δ Z) is (X, Y, correction Z); (X
Repair, Y
Repair, Z
Repair) be (X, Y, Z) revised theoretical coordinate.
With the artificial example of KUKA150-2 machine practical implementation step of the present invention is described below.
Step 1: in the envelope scope of industrial robot, be divided into a series of cube grid to the work space of treating of robot by given step-length 300mm.Be the simplified illustration problem, under basis coordinates system of robot, select x to coordinate from 1100 to 1400 here, y is to coordinate from 500 to 2000, and the zone of z to coordinate from 950 to 1850 is zone to be processed;
Step 2: measure and set up basis coordinates system of robot through laser tracker; Theoretical coordinate with eight summits of each cube grid of division in the step 1 positions in 19 ° of environment temperatures, 23 °, 26 °, 29 ° following control robot respectively, and measures and write down corresponding actual location coordinate with laser tracker;
Step 3: set up neural network model, and train with the test figure of gathering in the step 2.Through continuous test, confirm that finally the BP neural network forms by four layers, be respectively input layer, hidden layer 1, hidden layer 2, output layer, wherein input layer comprises 4 nodes, and 2 hidden layers are each self-contained 7 node all, and output layer comprises 3 nodes.The training function of network is ' trainlm ', and it adopts the Levenberg-Marquardt algorithm; The learning rate of network is 0.1; The sample number of network training is 480, and the sample number of checking is 5.The population number of confirming the PSO optimized Algorithm is 50, and the number of times of evolution is 600 times, and other relevant parameter is provided with as shown in the table.
480 groups of sample values that obtain with correlation parameter that is provided with and test are input in the program of matlab establishment and train, and the result is as shown in Figure 3, and the precision of prediction of 5 groups of test sample books of picked at random all below 0.1mm, satisfies accuracy requirement.In addition; For 475 groups of sample values that are used for network training, the training precision of most points surpasses the point of 0.1mm for minority below 0.1mm; Tracing it to its cause is because these points are in the edge of dividing grid, lacks the information of their characteristics of enough descriptions in the sample.And the numerical value below the numerical values recited of error and the repetitive positioning accuracy 0.15mm of robot is similar.
Step 4: in the zone of dividing to be processed, choose 5 points to be processed arbitrarily, its expectation coordinate figure is as shown in the table:
Regulating environment temperature changes temperature; Use coordinate figure in the table and combine the input of Current Temperatures as neural network; Actual location coordinate with eight summits of cube grid of living in that dope carries out the space lattice accuracy compensation, positions with revised coordinate figure control robot then, measures the actual location coordinate through laser tracker; Compare with the expectation coordinate with it, the result is as shown in the table:
Can know that through comparing result what the present invention proposed can significantly improve the absolute fix precision of industrial robot based on the industrial robot space lattice precision compensation method of particle group optimizing neural network.Advantage of the present invention is: 1) utilize particle swarm optimization algorithm to have the global optimization search and fast convergence rate gets characteristic, the initial weight of network is optimized, thereby it is slow to overcome the speed of convergence that the BP network algorithm exists, be absorbed in the shortcoming of local extremum easily; 2) optimum individual that provides with particle swarm optimization algorithm of BP algorithm utilizes the characteristics of its non-linear mapping and local optimal searching as the initial weight and the threshold value of network, carries out further optimizing, thereby obtains optimum network weight and threshold parameter; 3) volume coordinate information and ambient temperature information in the sample data of neural metwork training, have been comprised; Therefore compare traditional robot scaling method; The present invention had both considered the error that robot motion's mathematic(al) parameter, loading condition bring, and had also considered the error of bringing when environment temperature changes; 4) with certain step-length the enveloping space of robot being carried out space cube grid divides; For arbitrary target localization point that reaches; Predict eight summit actual location coordinates of the cube grid of the minimum that comprises this point through neural network; Combine the space lattice precision compensation method to carry out accuracy compensation to this impact point simply rapidly again, thereby can be implemented in the line compensation in industry spot.